多姿态人脸识别是模式识别领域的难题之一.首先采用水平镜像方法增加训练样本,并将所有训练样本在[-90,+90]的姿态范围内划分为7个子集,利用Gabor和2DPCA方法生成7个特征子空间.识别时,提取输入图像及其水平镜像图像的特征并分别向7个特征子空间投影,根据投影距离,采用决策融合判决策略得到最终的识别结果.在三个人脸库上的实验证明,在仅选取有限的多姿态训练样本的情况下,对旋转角度介于[-90,+90]的多姿态人脸能取得较高的识别率.%Multi-view face recognition is one of the posers in the field of pattern recognition. First, horizontal mirror image transformation is employed to increase the training samples, rrnand all samples are then divided into seven subsets within the pose range of [ - 90, + 90 ]. Secondly, Gabor wavelet and 2DPCA are used to generate seven feature subspaces. In recognition stage, the features of the inputted face images as well as their horizontal mirror images are extracted, and then to be projected onto seven feature subspaces, according to projection distance, the final recognition result can be achieved by utilising decision fusion judgement policy. Experiments based on three face libraries prove that in condition of only selecting limited multi-view training samples, it can obtain fairly high recognition rate on multi-view faces of which the rotation angles are between [ - 90, + 90 ] .
展开▼
